I think the title and description are at odds. I’m a software engineer, I check in tons of code, but I’m quickly getting to the point where I only look at the code, I don’t personally tap the keys to make the braces and expressions.

I’m a software engineer, but I’m not sure I “write” code by this market’s definition.

What does “write” mean in the description of this market?

@DavidFWatson
Comment by the question creator:
> Re: senior engineers who dont write code. As specified, software engineers are "the kind that write code, e.g. Python or C or JavaScript code", so they don't count.

I guess you could argue that this still leaves open the question of engineers that actively review code. It does say "write", but I can sympathize with the argument to count them.
@SanghyeonSeo

Or we'll just ask Chollet.

@SamuelKnoche to be clear, I’m not so senior that I’m just reviewing docs. I mean that I’m using AI so much that I rarely write the code myself, I read and review code that I asked the LLM to generate

@DavidFWatson I would argue that if you ever had to edit a small piece of AI generated code manually you still technically wrote code as part of your job

@Cactus but does the market creator agree? And what’s “had to” I can always edit the code by directing the ai to edit it.

@DavidFWatson that is true we should just wait for market creator to clarify

@DavidFWatson Yeah I was thinking they would just look up the number of people considered to be "software engineers" from the BLS or something. If there's going to be any attempt to reduce that number by some estimate of how many of them have become vibe coders, then the probability on this market should probably be a lot lower.

@DavidFWatson I think it’s very hard to measure which swes actually write code and it’s just cleaner to look up the numbers of reported swes. And if a senior SWE just reviews code and tells AI to iterate I would still consider that as a “coding” job

@Mochi I think it'd be okay if we count SWE that review code, though I don't know if that was the intention of the market creator. But if the job of SWE ends up evolving to become more like the one of a technical PM who only gives instructions to various LLMs without ever directly looking at a piece of code, then they should not be counted, even if they retain the title of SWE.

How many here tried tools like Cursor, Lovable, Windsurf, Replit, Bolt etc in the last month?

Tried using MCPs with above?

Keeping track of improvements in LLM capabilities?

25% of most recent YC batch let AI write >95% of their code.

I know software is eating the world, but AI will be writing the software.

@Moneyfolded

25% of most recent YC batch let AI write >95% of their code.

that's because they don't know how to code

Anyway, ai is still far from replacing human coders... far in terms of its capabilities. But not in terms of time. It'll be there within a year or two.

@jim

I'm curious, have you tried the tools I mentioned above in the last month? MCPs?

And do you think YC batch Software engineers are worse than the global average software engineer? Because that's approx what to compare with for this question.

@Moneyfolded I think the bottom 25% of YC founders are so braindead they don't even have the requisite moral agency to be termed "grifters".

Coding agents still just can't reliably do complicated / long time horizon tasks.

But this will change shortly. The time horizon of tasks LLMs can reliably do is supposedly increasing exponentially, roughly 5x-ing every year.

@Moneyfolded oh look this just got published! https://metr.org/hcast.pdf

> we find that current agentssucceed 70-80% of the time on tasks that take humans less than one hour,and less than 20% of the time on tasks that take humans more than 4 hours

@jim interesting read. Thanks!

Seen this one? It's similar to the estimate of 5x/year. Based on this, we're 4 years away from >100x length of tasks. With your data above, those <1h tasks with 70-80% success rate would be 1-3 work weeks.

I think it might become blurry who is a software engineer in a few years? Do they have to at least edit code or read code or something?

@Siebe I was assuming that

Prediction: there will be more software engineers (the kind that write code, e.g. Python or C or JavaScript code) in five years than there are today.

Implies that they have to actually be writing code. Not just "vibe coding" or directing agents or whatever

@jim ah yes, it's pretty clear in the description. I guess I just disagree with how high the market is then

@Siebe I think the figure will be less than 1% of the current number lol

Chollet's take is completely out of touch. The reality is, even today, years after LLM being a thing, most software engineers are glorified copy-paste artists, churning out boilerplate code day in and day out.

When is today?

@FranklinBaldo looks like the market was created 13 March 2024 - I agree, the description or the title should perhaps be updated to make it clear at a glance what the start date was/what "today" refers to

US or globally?

@LucaMasters Globally.

@SanghyeonSeo do you have a source or resolution process in mind?

@Tyler31 My current plan is to ask François Chollet.

@SanghyeonSeo ... and why would he know? this is a global census type question, that's just a guy